P ersonalized D iapause: Reducing Radio Energy Consumption of Smartphones by Network-Context Aware Dormancy Predictions 2012 Workshop on Power-Aware Computing and Systems October 7, 2012 1 Yeseong Kim and Jihong Kim Computer Architecture & Embedded Systems Lab. Department of Computer Science and Engineering Seoul National University A period of suspended growth accompanied by decreased metabolism in insects
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Personalized Diapause: Reducing Radio Energy Consumption of Smartphones
by Network-Context Aware Dormancy Predictions
2012 Workshop on Power-Aware Computing and Systems
October 7, 2012
1
Yeseong Kim and Jihong Kim
Computer Architecture & Embedded Systems Lab.
Department of Computer Science and Engineering
Seoul National University
A period of suspended growth
accompanied by decreased metabolism in insects
Radio Energy Consumption in Smartphones
• High radio energy consumption
• About 30% of the total energy consumption in smartphones (3G Network)
• College students, graduate students, bankers, kindergarten teachers …
• Study Period: during two weeks
• Method: using a modified Dalvik VM
• For logging network contexts with call stack information
25
Network
Context
Time Tail time
Netw
ork
Tra
nsm
issi
on
Whether/When did
a next transmission occur
in the tail time?
?
Personalized Network Usage Tendency
• Strong personalized network usage tendency
26
Mr. Every10Seconds Prof. EveryHour
0
20
40
60
80
Us
er
1
Us
er
2
Us
er
3
Us
er
4
Us
er
5
Us
er
6
Us
er
7
Us
er
8
Us
er
9
Us
er
10
User
11
Us
er
12
User
13
Us
er
14
User
15
Us
er
16
Us
er
17
Us
er
18
Us
er
19
Us
er
20
Us
er
21
Us
er
22
Us
er
23
Us
er
24
Us
er
25Rate
of
occu
rren
ce o
f a n
ext
tran
sm
issio
n
in t
he t
ail t
ime (
%)
Large energy wasted
0
10
20
30
40
50
60
3 4 5 6 7 8 9 10 11 12 13 14 15
Rate
of
occu
rren
ce o
f a n
ext
tran
sm
issio
n (
%)
Time of occurrence of a next transmission (sec)
Week 1
Week 2
Skewed Transmission Distribution
• User’s behavior for transmissions is quite skewed.
27
Most of the first transmissions in the tail
happen within the first 6 seconds.
These persistent right-skewed distribution
can be exploited to apply the fast dormancy feature.
Transmission Characteristics
for Different Network Activities Per User • Different transmission characteristics for different network activities
• The transmission trend of each NCB persists over long time.
28
0
10
20
30
40
50
60
70
80
1 2 3 4
Rate
of
occu
rren
ce o
f a n
ext
tran
sm
issio
n
in t
he t
ail t
ime (
%)
Week 1 Week 2
Checking
system update
Browsing
a web page
Sending
a message
Fetching
new emails
Exploiting these persistent transmission trends over different
NCBs ,we can estimate transmissions in the tail time.
Step 2.
Estimating Transmission Trend of Network Activity
• We estimate when/whether a transmission will occur in the tail time
based on the skewed transmission distribution of each NCB.
29
A user’s network context blocks
Transmission Trend
Transmission Trend
Transmission Trend Transmission Trend
Transmission Trend
Network
Context 1
Network
Context 2
Network
Context 4
Network
Context 3
Network
Context 5
Network
Context 6
NCB 1
NCB 2
NCB 3 NCB 4 NCB 5
Predictive Dormancy Transmission Trend
Key steps of Personalized Diapause
30
Extracting
semantically equivalent
network activities +
Network Context Block
Network
Context 1
Network
Context 3
Step 1. Step 2.
Estimating
transmission trends
of network activities
Predictive dormancy analysis
Step 3.
Step 3. Predictive Dormancy Analysis
• To determine when to invoke the fast dormancy feature • Considering cost-benefit tradeoff
• Intuitively, choosing the best moment (ti) to invoke fast dormancy
• Consider tk's only where the probability of retransmissions after tk is less than a given upper bound threshold
• (See our paper for a detail description of the decision procedure)
31
Time
Tail Energy
Invoke fast dormancy at ti
Pow
er
Expected Energy Benefit
(Bi)
Canceled Benefit
Energy Cost
(Cj)
If a transmission occurs at tj (pj)
Over
head
- ( × ) Probability
(pj) Gain (Gi) =
𝒋
Architectural Overview of Personalized Diapause
32
Android Framework
Dalvik VM
Application
Call Stack Tracer
Network Context Block
Extractor
Tail Time
Power Model Dormancy Granter
Cost-Benefit
Analysis Engine
Immediate-Successor
Trainer
Personalized Network
Activity Predictor
Personalized Network
Activity Predictor
• The key steps are implemented
as additional modules to the Dalvik VM and Android framework.
Outline
• Introduction
• Overview of Personalized Diapause
• Key Steps of Personalized Diapause
• Extraction of Network Context
• Estimation of Network Transmission Trend
• Predictive Dormancy Analysis
• Experimental Results
• Conclusion
33
• Implemented the Personalized Diapause (PD) technique
on Nexus S Android reference smartphones
• Running Android 2.3 (Gingerbread)
• To Dalvik VM and Android framework
• Using the collected network transmission logs from 25 users
• A custom log replayer tool reproduced network contexts logs.
• A 3G energy simulator was used for energy consumption comparison.
Experimental Environment
34
Log Replayer
Nexus S
(Target Device)
3G Energy Simulator User Log Transmission &
Fast dormancy Log
Impact of PD on Energy Consumption Saving
• Energy saving of Personalized Diapause
• No-fast-dormancy support as a baseline
35
0
10
20
30
40
50
60
User 1 User 2 User 3 User 4 Mean
En
erg
y s
avin
g (
%)
Mean (25 users)
10% 15% 20%
Reconnection increase limit
On average
23% energy saving
with 10% of reconnection increase
Impact of NCB Classification Technique
• Comparison with Per-user PD
• Assuming that all network contexts are classified to a single NCB.
36
0
5
10
15
20
25
30
35
40
User 1 User 2 User 3 User 4 Mean
En
erg
y S
avin
g (
%)
PD Per-user PD
Mean (25 users)
The fine-grained NCB separation based on semantic differences is
important in achieving a high energy efficiency.
Very poor energy saving
Per-NCB
Per-user Per-user vs
Conclusions
• We presented a general-purpose automatic predictive dormancy technique, Personalized Diapause. • Optimizing the radio energy consumption of smartphones with the fast
dormancy feature
• Personalized Diapause takes advantages of personalized network context usage in deciding when to release a radio connection. • Based on an automatic extraction technique of meaningful network
activities
• Future work • Extend for other types of system optimizations using other useful